Geological disaster identification method based on physical constraint and attention enhanced gating mechanism

Abstract

The present application belongs to the technical field of geological disaster monitoring, and particularly relates to a geological disaster identification method based on physical constraint and attention enhancement gating mechanism. First, multi-source data of a target geological area is collected and preprocessed to generate a multi-source data set. Then, multi-class physical prior feature data is obtained through a physical prior generation unit, and a maximum principal stress direction unit vector field and a stress concentration field are output through a stress field calculation unit. Subsequently, a deep learning inference unit containing a vibration physical constraint module, a stress constraint module and an attention enhancement gating module is constructed. Finally, the multi-source data, physical prior features and stress-related fields are input, and the geological disaster class probability is output through feature extraction, vibration vulnerability field and direction consistency modulation quantity generation, feature scaling, attention screening and up-sampling. The present application can solve the problems of low recognition accuracy and insufficient physical consistency in the existing geological disaster identification technology under complex geological environment.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster monitoring technology, and in particular relates to a geological disaster identification method based on physical constraints and attention-enhanced gating mechanism. Background Technology

[0002] The monitoring and identification of geological hazards is an important research direction in the fields of geological engineering and remote sensing technology. In existing technologies, geological hazard identification methods based on remote sensing imagery mainly rely on deep learning semantic segmentation models. These models extract features such as surface optical characteristics, topography, and vegetation to automatically classify and segment hazards such as landslides, collapses, and debris flows. These methods generally employ architectures such as convolutional neural networks, encoder-decoder structures, or attention mechanisms, including UNet, DeepLab, and SegMAN, and achieve feature extraction and target recognition through end-to-end training on large-scale image data. On the other hand, traditional geological hazard monitoring systems often combine multi-source data such as accelerometers, ground acoustic monitoring, and topographic change measurements, using threshold judgment or statistical feature analysis for hazard identification and early warning. These methods have been applied to the identification of various hazard types, providing a technological foundation for regional hazard monitoring.

[0003] However, existing deep learning-based methods for geological hazard identification primarily rely on data-driven learning processes, lacking coupling constraints with geophysical laws. In complex geological environments, remote sensing image features are easily affected by factors such as illumination, surface morphology, and noise interference, leading to errors in model identification results. Simultaneously, existing network models do not incorporate physical constraints related to geomechanics and vibration propagation during training, making it difficult for the learned features to reflect the intrinsic physical mechanisms of geological processes, resulting in insufficient generalization ability. Especially in real-world monitoring scenarios, physical signals such as surface vibration, stress changes, and sliding trends are directly related to hazard occurrence, but traditional models cannot effectively utilize this information for constrained learning. This lack of physical consistency in existing technologies limits the reliable application of deep learning models in real geological environments. Summary of the Invention

[0004] The technical problem solved by this invention is to provide a geological hazard identification method based on physical constraints and attention-enhanced gating mechanism, so as to solve the problems of low identification accuracy and insufficient physical consistency in existing geological hazard identification technologies under complex geological environments.

[0005] The basic solution provided by this invention is a geological hazard identification method based on physical constraints and attention-enhanced gating mechanisms, comprising:

[0006] S1: Collect data from the target geological area Remote sensing image data, digital elevation model data, geological structure and lithology data, rainfall data, and vibration and ground acoustic array data are preprocessed to obtain image raster. Elevation grid Slope grid Curvature grid Geological lithology coding Rainfall grid Spectral energy grid As a multi-source dataset, it is input into the basic data management unit for storage;

[0007] S2: Based on the obtained multi-source dataset, input the physical prior generation unit to obtain vibration amplification proxy maps respectively. Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram They are organized into a multi-scale pyramid structure according to a unified format to generate and store physical prior feature data.

[0008] S3: Elevation grid Geological lithology coding The stress is input into the stress field calculation unit to construct a plane stress equilibrium model, outputs the stress tensor at each grid position, and performs eigenvalue decomposition on the stress tensor to generate a unit vector field in the direction of the maximum principal stress. and stress concentration field ;

[0009] S4: Using the encoder-decoder structure as the backbone and the multi-scale semantic segmentation network as the baseline, a deep learning inference unit is constructed; in the deep learning inference unit, a vibration physical constraint module and a stress constraint module are connected on the encoder side, and an attention enhancement gating module is embedded on the decoder side.

[0010] S5: Combine multi-source datasets, physical prior feature data, and the unit vector field of the maximum principal stress direction. and stress concentration field The data is input into the deep learning inference unit, where it is processed by a multi-scale semantic segmentation network to obtain multi-scale feature maps. The vibration physical constraint module receives physical prior feature data and spectral energy grids. A vibration vulnerability field is generated; the unit vector field of the maximum principal stress direction is received through the stress constraint module. and stress concentration field A directionally consistent modulation amount is generated, and the multi-scale feature map is scaled element-wise in terms of channel and spatial dimensions according to the vibration vulnerability field and the directionally consistent modulation amount to obtain the modulated coded feature map.

[0011] S6: The attention-enhanced gating module performs element-wise multiplication and upsampling on the modulated coded features on each decoder, and outputs the probability of geological disaster categories through a pixel-level classification head.

[0012] Furthermore, S1 includes:

[0013] S1-1: Acquire remote sensing image data of the target geological area based on aerial or satellite platforms, and perform radiometric calibration and geometric correction on the remote sensing image data to obtain image raster. ;

[0014] S1-2: Obtain digital elevation model data of the target geological area from the surveying and mapping system, and generate elevation raster. And according to the elevation grid Calculate slope grid and curvature grid ;

[0015] S1-3: Obtain geological structure and lithology data based on geological survey data of the target geological region, and map the lithology classification to geological lithology codes. ;

[0016] S1-4: Acquire rainfall data based on the meteorological system of the target geological area, and then rasterize it into a rainfall raster after time averaging. ;

[0017] S1-5: Based on the accelerometer and ground acoustic sensor deployed in the target geological area, vibration and ground acoustic array signals are collected and processed sequentially, including mean removal and bandpass filtering, time alignment and window function slicing, calculation of short-time Fourier transform amplitude, frequency band integration and spatial interpolation, to generate a spectral energy grid. ;

[0018] S1-6: Store all raster data in the basic data management unit with the same spatial resolution, coordinate reference system, and storage format.

[0019] Furthermore, S2 includes:

[0020] S2-1: Based on the input slope grid Curvature grid Spectral Energy Raster Calculate and generate vibration amplification proxy diagram The expression is:

[0021]

[0022] in, , , This is the proportionality coefficient, and its value is greater than or equal to 0. Raster coordinates; Minimum-maximum normalization to interval The vibration amplification proxy diagram is obtained by renaming. ;

[0023] S2-2: Via elevation grid The gradient calculation yields the direction factor map. The expression is:

[0024]

[0025] in, This represents the first partial derivative of the digital elevation D in the direction x. This represents the first-order partial derivative of the digital elevation D with respect to the direction y. and Calculated using the central difference method. To prevent division by zero of constants; For the direction vector field, After normalization and renaming, the direction factor map is obtained. ;

[0026] S2-3: Based on geological lithology coding The local resonance frequency is calculated using the mapped shear wave velocity field and the formation thickness estimation function, expressed as follows:

[0027]

[0028] in, This is a function for estimating formation thickness. For shear wave velocity field, For local resonant frequencies; Perform logarithmic compression:

[0029]

[0030] Will Normalization to The renaming yields the resonant frequency proxy diagram. ;

[0031] S2-4: Based on the shear wave velocity field The spatial gradient strength is calculated using the following expression:

[0032]

[0033]

[0034] in, Let be the gradient vector of the shear wave velocity field. Shear wave velocity The partial derivative in the x-direction, Shear wave velocity Partial derivative in the y-direction; For gradient intensity; Perform linear normalization and rename to stiffness gradient index plot ;

[0035] S2-5: Vibration amplification proxy diagram obtained above Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram Spatial smoothing is performed, and after processing, the structures are organized into a multi-scale pyramid structure in a unified format and stored in a physical prior generation unit.

[0036] Furthermore, S3 includes:

[0037] S3-1: Elevation grid Geological lithology coding The stress field calculation unit is used to construct a plane stress equilibrium model, where any location within the target geological region is set. The displacement components in the x and y directions are , Then the strain tensor The expression is:

[0038]

[0039] in, For strain tensor, Let x be the linear strain in the x-direction. The linear strain is in the y-direction. Shear strain;

[0040] S3-2: Construct the plane stress constitutive equation based on the strain tensor, the expression of which is:

[0041]

[0042]

[0043]

[0044] in, For elastic modulus, Poisson's ratio, Let be the normal stress in the x-direction. The normal stress is in the y-direction. The shear stress is in the xy plane;

[0045] S3-3: Calculate the unit vector corresponding to the direction of the maximum principal stress, the expression is:

[0046]

[0047]

[0048] in, To prevent constants with zero denominators, the direction unit vector is used. Output in component form;

[0049] S3-4: Define the stress concentration field according to the degree of stress concentration. The expression is:

[0050]

[0051]

[0052]

[0053] in, To prevent constants with a denominator of zero, , For eigenvalues; This indicates the proportion of the local principal stress difference to the overall stress amplitude;

[0054] S3-5: Output direction unit vector and After spatial smoothing, the field is renamed as a unit vector field in the direction of maximum principal stress. and stress concentration field .

[0055] Furthermore, S4 includes:

[0056] S4-1: The encoder-decoder structure is used as the backbone, the baseline is a multi-scale semantic segmentation network, the pixel-level classification head outputs the classification probability, and a deep learning inference unit is constructed.

[0057] S4-2: The encoder consists of L cascaded scales, with the scale index as... The vibration physical constraint module and the stress constraint module are connected in parallel on the encoder side for outputting coded features.

[0058] S4-3: The decoder contains L upsampling stages, the... The first upsampling stage receives the upsampled output from the previous stage and applies it to the second upsampling stage through the upsampling operator. The spatial dimensions of each upsampling level are matched with the spatial dimensions of the encoded features to obtain... And concatenate the encoded features with the matching results to output the concatenated result. The expression is:

[0059]

[0060] in, The modulation code output by the stress constraint module;

[0061] S4-4: Embed the attention enhancement gating module after the concatenation of each upsampling stage of the decoder and before convolutional fusion.

[0062] Furthermore, S5 includes:

[0063] S5-1: Combine multi-source datasets, physical prior feature data, and the unit vector field of the maximum principal stress direction. and stress concentration field The input data is fed into the deep learning inference unit and assembled to obtain the input dataset. The expression is:

[0064]

[0065] S5-2: The input dataset is processed sequentially by the encoder. Downsampling and convolution are performed to obtain multi-scale feature maps. ;in Indicates the encoder at the 1st Level output characteristics;

[0066] S5-3: Integrate physical prior feature data and spectral energy raster Adjusted to match the multi-scale feature map using bilinear interpolation. The same space size, spliced ​​together according to the channel. Vibration physical constraint module for Convolution and normalization operations are performed to obtain the vibration vulnerability field. ; and the vibration vulnerability field Mapped to channel weight tensor , and multi-scale feature maps Vibration modulation to obtain coding features ;

[0067] S5-4: Unit vector field in the direction of maximum principal stress and stress concentration field Spatial dimensions and multi-scale feature maps Alignment, the stress constraint module will align the stress concentration field Channel weights are obtained through channel mapping. And based on unit vector fields Encoding features First-order anisotropic filtering is performed along the gradient direction to output encoded features. .

[0068] Furthermore, S5-3 specifically includes:

[0069] Physical prior feature data and spectral energy grid Generate input tensor The expression is:

[0070]

[0071] Among them, vibration amplification proxy diagram Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram Spectral energy grid Input tensor The dimensions after stacking through channels are H is the pixel height, and W is the pixel width. Number of pixel channels;

[0072] For input tensor Perform bilinear downsampling operation , so that it is consistent with the encoder in the first Feature maps generated at each scale By matching the spatial dimensions, we obtain:

[0073]

[0074] The vibration physical constraint module includes an input feature encoding layer, a physical consistency constraint layer, and a feature modulation layer, wherein the input feature encoding layer... Convolution and batch normalization:

[0075]

[0076]

[0077] The kernel sizes are respectively and The activation function is Batch normalization operation is ;

[0078] The physical consistency constraint layer is based on the results of convolution and batch normalization. Introducing the physical residual term, we obtain the discrete form of the residual, expressed as:

[0079]

[0080] in, For the residual discretization result, The local propagation velocity is jointly mapped by the stiffness gradient index map and the resonant frequency proxy map. , These are second-order difference operators along the x and y directions, respectively. It is a constant;

[0081] Discretization results of residuals Amplitude compression and standardization are performed to obtain the processed result. ;

[0082] The root of the feature modulation layer calculates the directional weights based on the directional factor field, which is expressed as follows:

[0083]

[0084] in, For the direction factor field, For the x-direction factor field, For the y-direction factor field;

[0085] The expression for calculating directional weights is:

[0086]

[0087] in, For directional weights;

[0088] The vibration vulnerability field is generated based on the vibration residual information, and its expression is:

[0089]

[0090] in, For the Sigmoid function, , , , , These are learnable parameters and are constants. Represents the vibration vulnerability field. For vibration amplification proxy field, For spectral energy field;

[0091] The vibration physical constraint module maps the vulnerable field through channel scaling. Generate the channel weight tensor, expressed as:

[0092]

[0093] in, , Indicates the total number of channels. Indicates the first One channel, Channel coefficient; Let the weight tensor of the c-th channel be used to combine the multi-scale feature map output by the encoder. The modulation feature is obtained by element-wise multiplication, and its expression is:

[0094]

[0095] in, This represents the modulation characteristics of the c-channel at the x-coordinate. Represents the multi-scale feature map of the x-coordinate and c-channel; modulated features Input into the stress constraint module.

[0096] Furthermore, S5-4 specifically includes:

[0097] Unit vector field in the direction of maximum principal stress and stress concentration field Bilinear downsampling operation Convert to:

[0098]

[0099]

[0100] in, The maximum principal stress is a unit vector field in the x-direction. The maximum principal stress is represented by a unit vector field in the y-direction.

[0101] The calculation process in the stress constraint module includes directional consistency constraints, stress modulation weight generation, and feature adjustment, among which:

[0102] The directional consistency constraint is used to measure the degree of deviation between the characteristic gradient and the principal stress directions. The expression for the characteristic gradient is:

[0103]

[0104] in, For the feature gradient, and The partial derivatives in the x and y directions are calculated respectively, and the deviation is calculated using the central difference discrete calculation.

[0105] Based on the unit vector field of the direction of maximum principal stress Define the directional consistency metric function:

[0106]

[0107] in, The function for measuring the directional consistency of the c-channel in the x-coordinate is... To prevent constants with a denominator of zero;

[0108] Stress modulation weights are generated based on the directional consistency metric function and the stress concentration field. The expression is as follows:

[0109]

[0110] in, This represents the stress modulation weight of the c-channel in the x-coordinate. It is a constant and a learnable coefficient. The stress concentration field represents the degree of stress concentration at the x-coordinate.

[0111] Apply stress modulation weights to the encoded features The modulated output is obtained, and the expression is:

[0112]

[0113] in, The modulated output of the c-channel for the x-coordinate;

[0114] Feature adjustment to modulate features Perform anisotropic smoothing once, and the output is linearly mapped to recover the feature scale:

[0115]

[0116] in, Linear convolution that preserves size.

[0117] Furthermore, S6 includes:

[0118] S6-1: The input tensor received by the attention-enhanced gating module includes the concatenation result. Vibration vulnerability field output by the vibration physical constraint module Scalar field output by the stress constraint module ;

[0119] S6-2: Calculate the gating weights based on the received input tensors, expressed as follows:

[0120]

[0121] in, Gating weights, range of values , For the Sigmoid function, for Convolution is mapped to a single-channel saliency extraction operator. Represents characteristic energy. Indicates the number of encoder channels. This represents the number of decoder-side feature channels concatenated with encoder features. , , For learnable scalar parameters;

[0122] The gating weights are modulated by element-wise multiplication to obtain:

[0123]

[0124] in, The modulated tensor; For the splicing result;

[0125] For the modulated tensor pass Fusion of convolutional and batch normalized layers:

[0126]

[0127] in For the first Layer decoding features;

[0128] By using pixel-level classification heads through Convolutional layers are mapped to class log-odds tensors :

[0129]

[0130] The pixel-level probability distribution is obtained through softmax operation:

[0131]

[0132] in, Let (i,j) be the probability that pixel (i,j) belongs to category k. To perform an exponential operation on the logarithmic odds of category k, This represents the summation of the nonnormalized probabilities over all K+1 classes;

[0133] Generate a classification map based on the maximum a posteriori principle:

[0134]

[0135] Output result: geological hazard segmentation mask , This represents the total number of geological disaster categories.

[0136] The principle and advantages of this invention are as follows: In the technical solution of this application, firstly, multi-source data such as remote sensing image data, digital elevation model data, geological structure and lithology data, rainfall data, vibration and ground acoustic array data of the target geological area are collected and preprocessed, and then input into the basic data management unit for storage; then, the physical prior generation unit generates physical prior feature data such as vibration amplification proxy map, direction factor map, resonance frequency proxy map, and stiffness gradient index map based on the multi-source dataset and organizes them into a multi-scale pyramid structure; then, the stress field calculation unit constructs a plane stress equilibrium model using elevation grid and geological lithology coding, outputs the stress tensor and decomposes it to obtain the unit vector field of the maximum principal stress direction and The stress concentration field is then used to construct a deep learning inference unit with an encoder-decoder backbone. The encoder side is connected to the vibration physics and stress constraint module, and the decoder side is embedded with the attention enhancement gating module. Finally, the multi-source dataset, physical prior feature data, and stress-related field are input into the inference unit. The multi-scale semantic segmentation network extracts features, the vibration physics constraint module generates the vibration vulnerability field, and the stress constraint module generates the directional consistency modulation amount. The two are scaled to obtain the modulated encoded feature map, which is then filtered and upsampled by the attention enhancement gating module. The probability of geological hazard category is output by the pixel-level classification head, realizing the accurate identification of geological hazards with "data-physics-mechanics" collaborative constraints.

[0137] The advantages are:

[0138] 1. Multi-source data fusion broadens the information dimension, and physical priors and stress fields introduce mechanistic knowledge, so that the features learned by the model not only contain statistical data regularities, but also conform to the physical and mechanical logic of geological disasters, thereby improving the ability of features to characterize disaster patterns.

[0139] 2. Physical and mechanical constraints provide "prior knowledge guidance" for the model, reducing ambiguity and erroneous learning in purely data-driven models. Attention gating focuses on key features, and multi-stage collaboration enables the model to more accurately identify disaster boundaries and classify categories in complex geological scenarios, resulting in stronger robustness.

[0140] 3. The explicit introduction of physical priors and stress fields links the model output with geological mechanisms, making it no longer a "black box"; it helps explain "why it was identified as this type of disaster", which meets the needs of mechanism analysis in geological disaster research and is also convenient for professionals to verify and apply.

[0141] 4. Multi-scale pyramid physical prior and multi-scale semantic segmentation network adapt to the multi-scale feature requirements of geological disasters from "local details to regional trends", effectively capture disaster features at different scales and improve the comprehensiveness of identification. Attached Figure Description

[0142] Figure 1 This is a general architecture diagram of an embodiment of the present invention;

[0143] Figure 2 This is the data acquisition and preprocessing process according to an embodiment of the present invention;

[0144] Figure 3 This is a schematic diagram illustrating the construction of a physical prior graph in an embodiment of the present invention;

[0145] Figure 4 This is a flowchart illustrating the calculation of the stress tensor field and direction according to an embodiment of the present invention.

[0146] Figure 5 This is a schematic diagram of the deep learning backbone and module embedding in an embodiment of the present invention;

[0147] Figure 6 This is a schematic diagram of the vibration physical constraint module processing in an embodiment of the present invention;

[0148] Figure 7 This is a schematic diagram of the stress constraint module processing according to an embodiment of the present invention;

[0149] Figure 8 This is a schematic diagram of the attention enhancement gating module processing in an embodiment of the present invention. Detailed Implementation

[0150] The following detailed description illustrates the specific implementation method:

[0151] The basic implementation examples are as follows: Figure 1 and Figure 2 As shown: A geological hazard identification method based on physical constraints and attention-enhanced gating mechanism, including:

[0152] S1: Collect data from the target geological area Remote sensing image data, digital elevation model data, geological structure and lithology data, rainfall data, and vibration and ground acoustic array data are preprocessed to obtain image raster. Elevation grid Slope grid Curvature grid Geological lithology coding Rainfall grid Spectral energy grid As a multi-source dataset, it is stored in the basic data management unit; among them, such as Figure 2 As shown, S1 includes:

[0153] S1-1: Acquire remote sensing image data of the target geological area based on aerial or satellite platforms, and perform radiometric calibration and geometric correction on the remote sensing image data to obtain image raster. In this process, radiometric calibration linearly normalizes pixel grayscale based on the sensor response function, while geometric correction uses affine or polynomial transformations with ground control points to align the image to a unified coordinate system. The corrected image is then cropped to the system's working area and invalid pixels are removed.

[0154] S1-2: Obtain digital elevation model data of the target geological area from the surveying and mapping system, and generate elevation raster. And according to the elevation grid Calculate slope grid and curvature grid The slope calculation formula is as follows:

[0155]

[0156] in, , The first-order partial derivatives of the digital elevation D in the x and y directions reflect the rate of change of elevation along the horizontal direction;

[0157] The formula for calculating curvature is:

[0158]

[0159] in, , The second-order partial derivatives of the digital elevation D in the x and y directions reflect the rate of change of elevation.

[0160] S1-3: Obtain geological structure and lithology data based on geological survey data of the target geological region, such as geological survey results or publicly available geological atlases, and map the lithology classification to geological lithology codes. ;

[0161] S1-4: Acquire rainfall data based on the meteorological system of the target geological area, and then rasterize it into a rainfall raster after time averaging. ;

[0162] S1-5: Based on the accelerometer and ground acoustic sensor deployed in the target geological area, vibration and ground acoustic array signals are collected and processed sequentially, including mean removal and bandpass filtering, time alignment and window function slicing, calculation of short-time Fourier transform amplitude, frequency band integration and spatial interpolation, to generate a spectral energy grid. ;

[0163] S1-6: Store all raster data in the basic data management unit with the same spatial resolution, coordinate reference system, and storage format; and provide a direct access interface for the subsequent physical prior generation unit, stress field calculation unit, and deep learning inference unit.

[0164] S2: Based on the obtained multi-source dataset, input the physical prior generation unit to obtain vibration amplification proxy maps respectively. Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram They are organized into a multi-scale pyramid structure according to a unified format to generate and store physical prior feature data; among them, such as Figure 3 As shown, S2 includes:

[0165] S2-1: Based on the input slope grid Curvature grid Spectral Energy Raster Calculate and generate vibration amplification proxy diagram The expression is:

[0166]

[0167] in, , , This is the proportionality coefficient, and its value is greater than or equal to 0. Raster coordinates; Minimum-maximum normalization to interval The vibration amplification proxy diagram is obtained by renaming. ;

[0168] S2-2: Via elevation grid The gradient calculation yields the direction factor map. The expression is:

[0169]

[0170] in, This represents the first partial derivative of the digital elevation D in the direction x. This represents the first-order partial derivative of the digital elevation D with respect to the direction y. and Calculated using the central difference method. To prevent division by zero constants; numerator The vector composed of elevation gradient components reflects the directional trend of terrain slope. The negative sign ensures that the direction vector points from high potential energy to low potential energy. For example, in landslide geological disasters, materials typically move from high altitude to low altitude, thus making the direction factor conform to actual physical laws. After normalization (modulus set to 1), the direction factor map is obtained by renaming. ;

[0171] S2-3: Based on geological lithology coding The local resonance frequency is calculated using the mapped shear wave velocity field and the formation thickness estimation function, expressed as follows:

[0172]

[0173] in, This is a function for estimating formation thickness. For shear wave velocity field, , , For local resonant frequencies; Perform logarithmic compression:

[0174]

[0175] Will Normalization to The renaming yields the resonant frequency proxy diagram. ;

[0176] S2-4: Based on the shear wave velocity field The spatial gradient strength is calculated using the following expression:

[0177]

[0178]

[0179] in, Let be the gradient vector of the shear wave velocity field. Shear wave velocity The partial derivative in the x-direction, Shear wave velocity Partial derivative in the y-direction; For gradient intensity; Perform linear normalization and rename to stiffness gradient index diagram ;

[0180] S2-5: Vibration amplification proxy diagram obtained above Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram Spatial smoothing is performed using a Gaussian filter kernel with radius r. After processing, the images are organized into a multi-scale pyramid structure in a unified format and stored in the physical prior generation unit. All generated prior images are output as floating-point raster files with spatial references and timestamps for use as input to the subsequent physical constraint module.

[0181] S3: Elevation grid Geological lithology coding The stress is input into the stress field calculation unit to construct a plane stress equilibrium model, outputs the stress tensor at each grid position, and performs eigenvalue decomposition on the stress tensor to generate a unit vector field in the direction of the maximum principal stress. and stress concentration field Among them, such as Figure 4 As shown, S3 includes:

[0182] S3-1: Elevation grid Geological lithology coding The stress field calculation unit is used to construct a plane stress equilibrium model, where any location within the target geological region is set. The displacement components in the x and y directions are , Then the strain tensor The expression is:

[0183]

[0184] in, For strain tensor, Let x be the linear strain in the x-direction. The linear strain is in the y-direction. Shear strain;

[0185] S3-2: Construct the plane stress constitutive equation based on the strain tensor, the expression of which is:

[0186]

[0187]

[0188]

[0189] in, For elastic modulus, Poisson's ratio, Let be the normal stress in the x-direction. The normal stress is in the y-direction. The shear stress is in the xy plane;

[0190] S3-3: Calculate the unit vector corresponding to the direction of maximum principal stress. First, define the target geological area. Discretized into a regular grid with a step size of , The emergent difference scheme is used to discretely evaluate the scoring equation, with nodes... Indicates position ,but:

[0191]

[0192]

[0193] in, for The central difference approximation of the partial derivatives with respect to x for The central difference approximation of the partial derivatives with respect to y; , , , All are nodes;

[0194] right and Perform the same approximate calculations as above to form a system of linear equations, and then use iterative solution methods, such as the conjugate gradient method, to calculate the values ​​at the nodes. , , The value of is obtained by solving for the stress tensor field:

[0195]

[0196] The expression for the unit vector corresponding to the direction of maximum principal stress is:

[0197]

[0198]

[0199] in, To prevent constants with zero denominators, the direction unit vector is used. Output in component form;

[0200] S3-4: Define the stress concentration field according to the degree of stress concentration. The expression is:

[0201]

[0202]

[0203]

[0204] in, To prevent constants with a denominator of zero, , For eigenvalues; This indicates the proportion of the local principal stress difference to the overall stress amplitude;

[0205] S3-5: Output direction unit vector and After spatial smoothing, the field is renamed as a unit vector field in the direction of maximum principal stress. and stress concentration field .

[0206] S4: A deep learning inference unit is constructed using an encoder-decoder structure as the backbone and a multi-scale semantic segmentation network as the baseline. In this deep learning inference unit, a vibration physical constraint module and a stress constraint module are integrated on the encoder side, and an attention-enhanced gating module is embedded on the decoder side. Wherein, as... Figure 5 As shown, S4 includes:

[0207] S4-1: The encoder-decoder structure is used as the backbone, the baseline is a multi-scale semantic segmentation network, the pixel-level classification head outputs the classification probability, and a deep learning inference unit is constructed.

[0208] S4-2: The encoder consists of L cascaded scales, with the scale index as... In all subsequent content of this article, the upper right corner All indicate the first Level Scale Index; The spatial size of the level feature map is The number of channels is ,in and For pixel height and width, the vibration physical constraint module and the stress constraint module are connected in parallel on the encoder side for outputting encoded features;

[0209] S4-3: The decoder contains L upsampling stages, the... Each upsampling stage receives the upsampling output from the previous stage. In the decoder, the numbers are numbered from top to bottom, corresponding one-to-one with the encoding level. hour Defined as the result of a linear transformation of the top-level encoded features; and then applied to the th... The spatial dimensions of each upsampling level are matched with the spatial dimensions of the encoded features to obtain... And concatenate the encoded features with the matching results to output the concatenated result. The expression is:

[0210]

[0211] in, The modulation code output by the stress constraint module;

[0212] S4-4: Embed the attention enhancement gating module after the concatenation of each upsampling stage of the decoder and before convolutional fusion.

[0213] S5: Combine multi-source datasets, physical prior feature data, and the unit vector field of the maximum principal stress direction. and stress concentration field The data is input into the deep learning inference unit, where it is processed by a multi-scale semantic segmentation network to obtain multi-scale feature maps. The vibration physical constraint module receives prior physical feature data and spectral energy grids. A vibration vulnerability field is generated; the unit vector field of the maximum principal stress direction is received through the stress constraint module. and stress concentration field A directionally consistent modulation quantity is generated, and the multi-scale feature map is element-wise scaled in terms of channel and spatial dimensions based on the vibration vulnerability field and the directionally consistent modulation quantity to obtain the modulated coded feature map; where, for example Figure 5 As shown, S5 includes:

[0214] S5-1: Combine multi-source datasets, physical prior feature data, and the unit vector field of the maximum principal stress direction. and stress concentration field The input data is fed into the deep learning inference unit and assembled to obtain the input dataset. The expression is:

[0215]

[0216] S5-2: The input dataset is processed sequentially by the encoder. Downsampling and convolution are performed to obtain multi-scale feature maps. ;in Indicates the encoder at the 1st Level output characteristics;

[0217] S5-3: Integrate physical prior feature data and spectral energy raster Adjusted to match the multi-scale feature map using bilinear interpolation. The same space size, spliced ​​together according to the channel. Vibration physical constraint module for Convolution and normalization operations are performed to obtain the vibration vulnerability field. ; and the vibration vulnerability field Mapped to channel weight tensor , and multi-scale feature maps Vibration modulation to obtain coding features Specifically:

[0218] Physical prior feature data and spectral energy grid Generate input tensor The expression is:

[0219]

[0220] Among them, vibration amplification proxy diagram Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram Spectral energy grid Input tensor The dimensions after stacking through channels are H is the pixel height, and W is the pixel width. Number of pixel channels;

[0221] The encoder in the The feature maps generated at each scale are , , For the first The pixel height and width at each scale, For the first The number of channels per scale; for the input tensor Perform bilinear downsampling operation , so that it is consistent with the encoder in the first Feature maps generated at each scale By matching the spatial dimensions, we obtain:

[0222]

[0223] The vibration physical constraint module includes an input feature encoding layer, a physical consistency constraint layer, and a feature modulation layer, such as... Figure 6 As shown, the input feature encoding layer pairs Convolution and batch normalization:

[0224]

[0225]

[0226] The kernel sizes are respectively and The activation function is Batch normalization operation is ;

[0227] The physical consistency constraint layer is based on the results of convolution and batch normalization. Introducing the physical residual term, we obtain the discrete form of the residual, expressed as:

[0228]

[0229] in, For the residual discretization result, The local propagation velocity is jointly mapped by the stiffness gradient index map and the resonant frequency proxy map. , These are second-order difference operators along the x and y directions, respectively. It is a constant;

[0230] Discretization results of residuals Amplitude compression and standardization are performed to obtain the processed result. :

[0231]

[0232] The root of the feature modulation layer calculates the directional weights based on the directional factor field, which is expressed as follows:

[0233]

[0234] in, For the direction factor field, For the x-direction factor field, For the y-direction factor field;

[0235] The expression for calculating directional weights is:

[0236]

[0237] in, For directional weights;

[0238] The vibration vulnerability field is generated based on the vibration residual information, and its expression is:

[0239]

[0240] in, For the Sigmoid function, , , , , These are learnable parameters and are constants. Represents the vibration vulnerability field. For vibration amplification proxy field, For spectral energy field;

[0241] The vibration physical constraint module maps the vulnerable field through channel scaling. Generate the channel weight tensor, expressed as:

[0242]

[0243] in, , Indicates the total number of channels. Indicates the first One channel, Channel coefficient; Let the weight tensor of the c-th channel be used to combine the multi-scale feature map output by the encoder. The modulation feature is obtained by element-wise multiplication, and its expression is:

[0244]

[0245] in, This represents the modulation characteristics of the c-channel at the x-coordinate. Represents the multi-scale feature map of the x-coordinate and c-channel; modulated features Input into the stress constraint module.

[0246] Therefore, in the above calculation process, each operator of the vibration physics constraint module maintains the same spatial dimension, and the input tensor... With output tensor , and There is a one-to-one correspondence between them, module parameters , , , , , The module can learn during training, and its internal calculations do not involve random sampling or probabilistic processes, ensuring the determinism and reproducibility of the output results.

[0247] S5-4: Unit vector field in the direction of maximum principal stress and stress concentration field Spatial dimensions and multi-scale feature maps Alignment, the stress constraint module will align the stress concentration field Channel weights are obtained through channel mapping. And based on unit vector fields Encoding features First-order anisotropic filtering is performed along the gradient direction to output encoded features. Specifically:

[0248] Unit vector field in the direction of maximum principal stress and stress concentration field Bilinear downsampling operation Convert to:

[0249]

[0250]

[0251] in, The maximum principal stress is a unit vector field in the x-direction. The maximum principal stress is represented by a unit vector field in the y-direction.

[0252] The calculation process in the stress constraint module includes directional consistency constraints, stress modulation weight generation, and feature adjustment, such as... Figure 7 As shown, where:

[0253] The directional consistency constraint is used to measure the degree of deviation between the characteristic gradient and the principal stress directions. The expression for the characteristic gradient is:

[0254]

[0255] in, For the feature gradient, and The partial derivatives in the x and y directions are calculated respectively, and the deviation is calculated using the central difference discrete calculation.

[0256] Based on the unit vector field of the direction of maximum principal stress Define the directional consistency metric function:

[0257]

[0258] in, The function for measuring the directional consistency of the c-channel in the x-coordinate is... To prevent constants with a denominator of zero;

[0259] Stress modulation weights are generated based on the directional consistency metric function and the stress concentration field. The expression is as follows:

[0260]

[0261] in, This represents the stress modulation weight of the c-channel in the x-coordinate. It is a constant and a learnable coefficient. This represents the stress concentration level of the stress concentration field at the x-coordinate; if the characteristic gradient direction coincides with the principal stress direction, then... Greater than 1; if the direction is opposite, then Less than 1, thus forming physical modulation based on stress direction in the convolutional features;

[0262] Apply stress modulation weights to the encoded features The modulated output is obtained, and the expression is:

[0263]

[0264] in, The modulated output of the c-channel for the x-coordinate;

[0265] Feature adjustment to modulate features Perform anisotropic smoothing once, and the output is linearly mapped to recover the feature scale:

[0266]

[0267] in, Linear convolution that preserves size.

[0268] In the aforementioned vibration physical constraint module and stress constraint module, the two modules are used in series. The stress constraint module acts on the feature map modulated by the vibration physical constraint module, and its output is directly used by the decoding stage or the attention enhancement gating module to achieve consistent constraint between the direction of ground stress and the depth feature space.

[0269] S6: The attention-enhanced gating module performs element-wise multiplication filtering and upsampling on the modulated coded features across each decoder, and outputs the probability of geological disaster categories through a pixel-level classification head; among which, such as Figure 8 As shown, S6 includes:

[0270] S6-1: The input tensor received by the attention-enhanced gating module includes the concatenation result. Vibration vulnerability field output by the vibration physical constraint module Scalar field output by the stress constraint module ;

[0271] S6-2: Calculate the gating weights based on the received input tensor. First, perform channel attention calculation. Perform global average pooling and global max pooling along the spatial dimension to obtain two sets of channel description vectors:

[0272]

[0273]

[0274] in, , For channel indexing, This refers to the number of splicing channels; For global average pooling, , For height and width indices in the spatial dimension, Let l be the total number of spatial pixels in the feature map of layer l. This is for global max pooling;

[0275] Global average pooling and global max pooling Shared fully connected mapping And summed and then passed through the Sigmoid function The channel weights are obtained upon activation, expressed as follows:

[0276]

[0277] For channel weights and for matching space dimensions, we expand them as follows:

[0278]

[0279] in, The channel weights are those expanded along spatial dimensions i and j.

[0280] Next, spatial attention calculation is performed. Perform average pooling and max pooling along the channel dimension, as shown in the expression:

[0281]

[0282]

[0283] in, , For height and width indices in the spatial dimension, To concatenate the number of channels, For average pooling, For max pooling;

[0284] After splicing Convolution and Sigmoid activation yield a spatial attention map:

[0285]

[0286] in, ;

[0287] Finally, the gating weight is calculated, expressed as follows:

[0288]

[0289] in, Gating weights, range of values This represents the activation ratio at each pixel location. For the Sigmoid function, for Convolution is mapped to a single-channel saliency extraction operator. Represents characteristic energy. Indicates the number of encoder channels. This represents the number of decoder-side feature channels concatenated with encoder features. , , For learnable scalar parameters;

[0290] The gating weights are modulated by element-wise multiplication to obtain:

[0291]

[0292] in, The modulated tensor; specifically:

[0293]

[0294] For the modulated tensor pass Fusion of convolutional and batch normalized layers:

[0295]

[0296] in For the first Layer decoding features;

[0297] By using pixel-level classification heads through Convolutional layers are mapped to class log-odds tensors :

[0298]

[0299] The pixel-level probability distribution is obtained through softmax operation:

[0300]

[0301] in, Let (i,j) be the probability that pixel (i,j) belongs to category k. To perform an exponential operation on the logarithmic odds of category k, This represents the summation of the nonnormalized probabilities over all K+1 classes;

[0302] Generate a classification map based on the maximum a posteriori principle:

[0303]

[0304] Output result: geological hazard segmentation mask , This represents the total number of geological disaster categories.

[0305] To ensure that the output results are consistent with geospatial data, the following post-processing steps are performed:

[0306] 1. Spatial smoothing: for conduct Neighborhood majority filtering eliminates isolated noise pixels.

[0307] 2. Regional connectivity correction: A connected component labeling algorithm is used to identify independent disaster areas, and areas smaller than the area threshold are deleted. An isolated fragment.

[0308] 3. Geographic projection restoration: Based on the spatial reference information of the input data, the output mask is mapped back to the original coordinate system, and a geographic reference file is output.

[0309] 4. Disaster Category Probability Map Output: In addition to the final classification results, the system also outputs a category probability distribution map for subsequent risk assessment or uncertainty analysis. The output includes:

[0310] Geological Disaster Classification Mask Probability distribution diagrams for each category Vibration and stress constraint response diagrams Final inference log and parameter snapshot files.

[0311] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A geological hazard identification method based on physical constraints and attention-enhanced gating mechanism, characterized in that: include: S1: Collect data from the target geological area Remote sensing image data, digital elevation model data, geological structure and lithology data, rainfall data, and vibration and ground acoustic array data are preprocessed to obtain image raster. Elevation grid Slope grid Curvature grid Geological lithology coding Rainfall grid Spectral energy grid As a multi-source dataset, it is input into the basic data management unit for storage; S2: Based on the obtained multi-source dataset, input the physical prior generation unit to obtain vibration amplification proxy maps respectively. Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram They are organized into a multi-scale pyramid structure according to a unified format to generate and store physical prior feature data. S3: Elevation grid Geological lithology coding The stress is input into the stress field calculation unit to construct a plane stress equilibrium model, outputs the stress tensor at each grid position, and performs eigenvalue decomposition on the stress tensor to generate a unit vector field in the direction of the maximum principal stress. and stress concentration field ; S4: Using the encoder-decoder structure as the backbone and the multi-scale semantic segmentation network as the baseline, a deep learning inference unit is constructed; in the deep learning inference unit, a vibration physical constraint module and a stress constraint module are connected on the encoder side, and an attention enhancement gating module is embedded on the decoder side. S5: Combine multi-source datasets, physical prior feature data, and the unit vector field of the maximum principal stress direction. and stress concentration field The data is input into the deep learning inference unit, where it is processed by a multi-scale semantic segmentation network to obtain multi-scale feature maps. The vibration physical constraint module receives physical prior feature data and spectral energy grids. A vibration vulnerability field is generated; the unit vector field of the maximum principal stress direction is received through the stress constraint module. and stress concentration field A directionally consistent modulation amount is generated, and the multi-scale feature map is scaled element-wise in terms of channel and spatial dimensions according to the vibration vulnerability field and the directionally consistent modulation amount to obtain the modulated coded feature map. S6: The attention-enhanced gating module performs element-wise multiplication and upsampling on the modulated coded features on each decoder, and outputs the probability of geological disaster categories through a pixel-level classification head.

2. The method of claim 1, wherein the method is based on a physical constraint and attention enhanced gating mechanism. S1 includes: S1-1: Obtain remote sensing image data of a target geological region based on an aerial or satellite platform, and perform radiometric calibration and geometric correction processing on the remote sensing image data to obtain image grids ; S1-2: Obtain digital elevation model data of the target geological area from the surveying and mapping system, and generate elevation raster. And according to the elevation grid Calculate slope grid and curvature grid ; S1-3: Obtain geological structure and lithology data based on the geological survey data of the target geological area, and map the classification of the lithology to geological lithology codes through coding ; S1-4: Obtain rainfall data based on the meteorological system of the target geological area, and rasterize the time-averaged rainfall data into a rainfall grid ; S1-5: Based on the accelerometer and ground acoustic sensor deployed in the target geological area, vibration and ground acoustic array signals are collected and processed sequentially through mean removal and bandpass filtering, time alignment and window function slicing, short-time Fourier transform amplitude calculation, frequency band integration and spatial interpolation to generate a spectral energy grid. ; S1-6: Store all raster data in the basic data management unit with the same spatial resolution, coordinate reference system, and storage format.

3. The method of claim 1, wherein the method is based on a physical constraint and attention enhanced gating mechanism. S2 includes: S2-1: Based on the input slope grid Curvature grid Spectral Energy Raster Calculate and generate vibration amplification proxy diagram The expression is: in, , , This is the proportionality coefficient, and its value is greater than or equal to 0. Raster coordinates; Minimum-maximum normalization to interval The vibration amplification proxy diagram is obtained by renaming. ; S2-2: Direction factor map is calculated by the gradient of the elevation grid , with the expression:​ in, This represents the first partial derivative of the digital elevation D in the direction x. This represents the first-order partial derivative of the digital elevation D with respect to the direction y. and Calculated using the central difference method. To prevent division by zero of constants; For the direction vector field, After normalization and renaming, the direction factor map is obtained. ; S2-3: According to the geological lithology code The mapped shear wave velocity field and the formation thickness estimate function compute the local resonant frequency, expressed as: wherein, is a formation thickness estimation function, is a shear wave velocity field, is a local resonant frequency; the local resonant frequency is log-compressed: Will Normalization to The renaming yields the resonant frequency proxy diagram. ; S2-4: From the shear wave velocity field The spatial gradient strength is calculated as: in, Let be the gradient vector of the shear wave velocity field. Shear wave velocity The partial derivative in the x-direction, Shear wave velocity Partial derivative in the y-direction; For gradient intensity; Perform linear normalization and rename to stiffness gradient index diagram ; S2-5: Vibration amplification proxy diagram obtained above Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram Spatial smoothing is performed, and after processing, the structures are organized into a multi-scale pyramid structure in a unified format and stored in a physical prior generation unit.

4. The method of claim 1, wherein the method is based on a physical constraint and attention enhanced gating mechanism. S3 includes: S3-1: Elevation grid Geological lithology coding The stress field calculation unit is used to construct a plane stress equilibrium model, where any location within the target geological region is set. The displacement components in the x and y directions are , Then the strain tensor The expression is: wherein, is the strain tensor, is the linear strain in the x direction, is the linear strain in the y direction, is the shear strain; S3-2: Construct the plane stress constitutive equation based on the strain tensor, the expression of which is: in, For elastic modulus, Poisson's ratio, Let be the normal stress in the x-direction. The normal stress is in the y-direction. The shear stress is in the xy plane; S3-3: Calculate the unit vector corresponding to the direction of the maximum principal stress, the expression is: in, This is the unit vector corresponding to the direction of maximum principal stress. To prevent constants with zero denominators, the direction unit vector is used. Output in component form; S3-4: Define the stress concentration field according to the degree of stress concentration. The expression is: wherein to prevent the denominator from being zero, , is the eigenvalue; denotes the proportion of the local principal stress difference to the total stress amplitude; S3-5: Output direction unit vector and After spatial smoothing, the field is renamed as a unit vector field in the direction of maximum principal stress. and stress concentration field .

5. The geological hazard identification method based on physical constraints and attention-enhanced gating mechanism according to claim 1, characterized in that: S4 includes: S4-1: The encoder-decoder structure is used as the backbone, the baseline is a multi-scale semantic segmentation network, the pixel-level classification head outputs the classification probability, and a deep learning inference unit is constructed. S4-2: the encoder consists of L scale cascades, and the scale index is ; the vibration physical constraint module and the stress constraint module are connected in parallel on the encoder side and used for outputting the encoding features; S4-3: The decoder contains L upsampling stages, the... The first upsampling stage receives the upsampled output from the previous stage and applies it to the second upsampling stage through the upsampling operator. The spatial dimensions of each upsampling level are matched with the spatial dimensions of the encoded features to obtain... And concatenate the encoded features with the matching results to output the concatenated result. The expression is: wherein, modulation encoding output by the stress-constrained module; S4-4: Embed the attention enhancement gating module after the concatenation of each upsampling stage of the decoder and before convolutional fusion.

6. The method of claim 5, wherein the method is based on a physical constraint and attention enhanced gating mechanism for geological disaster identification. S5 includes: S5-1: inputting the multi-source data set, the physical prior feature data, the maximum principal stress direction unit vector field and the stress concentration field into a deep learning inference unit, and assembling to obtain an input data set , and the expression is: S5-2: The input dataset is processed sequentially by the encoder. Downsampling and convolution are performed to obtain multi-scale feature maps. ;in Indicates the encoder at the 1st Level output characteristics; S5-3: Integrate physical prior feature data and spectral energy raster Adjusted to match the multi-scale feature map using bilinear interpolation. The same space size, spliced ​​together according to the channel. Vibration physical constraint module for Convolution and normalization operations are performed to obtain the vibration vulnerability field. ; and the vibration vulnerability field Mapped to channel weight tensor , and multi-scale feature maps Vibration modulation to obtain coding features ; S5-4: Unit vector field in the direction of maximum principal stress and stress concentration field Spatial dimensions and multi-scale feature maps Alignment, the stress constraint module will align the stress concentration field Channel weights are obtained through channel mapping. And based on unit vector fields Encoding features First-order anisotropic filtering is performed along the gradient direction to output encoded features. .

7. The method of claim 6, wherein the method is based on a physical constraint and attention enhanced gating mechanism for geological disaster identification. Specifically, S5-3 is as follows: combining physical prior feature data and spectral energy grid generating an input tensor , expressed as: Among them, vibration amplification proxy diagram Direction factor diagram Resonance frequency proxy diagram Stiffness gradient index diagram Spectral energy grid Input tensor The dimensions after stacking through channels are H is the pixel height, and W is the pixel width. Number of pixel channels; For input tensor Perform bilinear downsampling operation , so that it is consistent with the encoder in the first Feature maps generated at each scale By matching the spatial dimensions, we obtain: wherein, is an input tensor performing a bilinear down-sampling operation outputting a result; The vibration physical constraint module includes an input feature encoding layer, a physical consistency constraint layer, and a feature modulation layer, wherein the input feature encoding layer encodes the input feature Convolution and batch normalization processing: wherein, is the result of the first convolution and batch normalization processing, is the result of the second convolution and batch normalization processing, the convolution kernel size is and , the activation function is , and the batch normalization operation is ; Physical consistency constraint layer according to convolution and batch normalization processing results Introducing a physical residual term, the residual discrete form is obtained, and the expression is: wherein, is the residual discrete result, is the local propagation velocity jointly mapped by the stiffness gradient indicator map and the resonance frequency proxy map, , are second-order difference operators along the x and y directions, respectively, is a constant; Discretization results of residuals Amplitude compression and standardization are performed to obtain the processed result. ; The root of the feature modulation layer calculates the directional weights based on the directional factor field, which is expressed as follows: wherein is a field of directional factors, is a field of x-directional factors, is a field of y-directional factors; The expression for calculating directional weights is: wherein is a directional weight; The vibration vulnerability field is generated based on the vibration residual information, and its expression is: wherein, is a Sigmoid function, , , , , are learnable parameters, belonging to constants; denotes a vibrational vulnerability field, is a vibrational amplification proxy field, is a spectral energy field; The vibration physics constraint module maps the vulnerability field through a channel scaling map A channel weight tensor is generated, expressed as: in, , Indicates the total number of channels. Indicates the first One channel, Channel coefficient; Let the weight tensor of the c-th channel be used to combine the multi-scale feature map output by the encoder. The modulation feature is obtained by element-wise multiplication, and its expression is: in, This represents the modulation characteristics of the c-channel at the x-coordinate. Represents the multi-scale feature map of the x-coordinate and c-channel; modulated features Input into the stress constraint module.

8. The geological hazard identification method based on physical constraints and attention-enhanced gating mechanism according to claim 7, characterized in that: Specifically, S5-4 is: Unit vector field in the direction of maximum principal stress and stress concentration field Bilinear downsampling operation Convert to: in, The maximum principal stress is a unit vector field in the x-direction. The maximum principal stress is represented by a unit vector field in the y-direction. This is the stress concentration field result obtained after downsampling. This is the result of the unit vector field in the direction of maximum principal stress obtained after downsampling; The calculation process in the stress constraint module includes directional consistency constraints, stress modulation weight generation, and feature adjustment, among which: The directional consistency constraint is used to measure the degree of deviation between the characteristic gradient and the principal stress directions. The expression for the characteristic gradient is: in, For the feature gradient, and The partial derivatives in the x and y directions are calculated respectively, and the deviation is calculated using the central difference discrete calculation. Based on the unit vector field of the direction of maximum principal stress Define the directional consistency metric function: in, The function for measuring the directional consistency of the c-channel in the x-coordinate. To prevent constants with a denominator of zero; This represents the unit vector field in the x-direction of the maximum principal stress after downsampling. This represents the unit vector field in the y-direction of the maximum principal stress after downsampling; Modulation characteristics Partial derivative in the x-direction at the x-coordinate and the c-channel. Modulation characteristics Partial derivative in the y direction at the x-coordinate and the c-channel; Stress modulation weights are generated based on the directional consistency metric function and the stress concentration field. The expression is as follows: in, This represents the stress modulation weight of the c-channel in the x-coordinate. It is a constant and a learnable coefficient. The stress concentration field is at the x-coordinate, representing the degree of stress concentration. Apply stress modulation weights to the encoded features The modulated output is obtained, and the expression is: in, The modulated output of the c-channel for the x-coordinate; Feature adjustment to modulate features Perform anisotropic smoothing once, and the output is linearly mapped to recover the feature scale: in, Linear convolution that preserves size.

9. The geological hazard identification method based on physical constraints and attention-enhanced gating mechanism according to claim 8, characterized in that: S6 includes: S6-1: The input tensor received by the attention-enhanced gating module includes the concatenation result. Vibration vulnerability field output by the vibration physical constraint module Scalar field output by the stress constraint module ; S6-2: Calculate the gating weights based on the received input tensors, expressed as follows: in, Gating weights, range of values , For the Sigmoid function, for Convolution is mapped to a single-channel saliency extraction operator. Represents characteristic energy. Indicates the number of encoder channels. This represents the number of decoder-side feature channels concatenated with encoder features. , , For learnable scalar parameters; The gating weights are modulated by element-wise multiplication to obtain: in, The modulated tensor; For the splicing result; For the modulated tensor pass Fusion of convolutional and batch normalized layers: in For the first Layer decoding features; By using pixel-level classification heads through Convolutional layers are mapped to class log-odds tensors : The pixel-level probability distribution is obtained through softmax operation: in, Let (i,j) be the probability that pixel (i,j) belongs to category k. To perform an exponential operation on the logarithmic odds of category k, This represents the summation of the nonnormalized probabilities over all K+1 classes; Generate a classification map based on the maximum a posteriori principle: Output result: geological hazard segmentation mask , This represents the total number of geological disaster categories.

Images